Silicon Valley VCs begin hoarding GPUs: a16z stockpiles 20,000 cards with plans to expand, competing for AI startups goes beyond just funding.

Beating AI News Flash: Multiple Silicon Valley venture capital firms are racing to lock up GPU computing power to provide exclusive resources for their AI portfolio companies. a16z established a computing pool of approximately 20,000 GPUs as early as 2024 and is now in talks with data center companies Switch and Volta for further expansion. Another VC firm, General Catalyst, plans to build data centers for its portfolio companies and in May of this year specifically recruited the former head of Nvidia's DGX Cloud to oversee its AI infrastructure business.

Why are VCs starting to personally solve the computing power problem? Because AI startups, even if they raise money, may not be able to rent GPUs in time. In April of this year, major cloud providers like Microsoft were reported to be prioritizing computing power for internal teams and large customers, leaving some startups facing months-long wait times. a16z previously disclosed that when computing power is scarce, startups seeking the price they could have gotten with a six-month contract may be asked to sign a three-year deal. But model training requires short-term concentrated use of large amounts of GPUs, and inference demand after product launch is difficult to predict. Locking in years of computing power in advance can easily lead to wasted capital.

a16z's approach is to centrally procure and lease computing power through a project called Oxygen, then distribute it to its portfolio companies, allowing them to enjoy the prices and contract terms available only to large customers. Now, this model is extending upstream. In August of this year, a16z co-led a $300 million funding round for Volta, which reached a valuation of $2.4 billion. It invests in both AI startups and data center companies that provide computing power, attempting to control both capital and GPU supply channels simultaneously.

Moreover, computing power itself is starting to become a business for VCs. Small VC firm Crucible Capital built a cluster with 576 Nvidia B300 GPUs for its portfolio companies to test products, while also selling some computing power to a frontier AI lab. This ensures the needs of portfolio companies while also turning idle computing power into revenue. For VCs, providing GPUs can also attract founders who need large amounts of computing power, increasing the chances of securing hot investment deals.

However, this model also makes VCs bear risks that were previously mainly borne by cloud providers and data center operators. GPUs iterate quickly, and long-term leasing or building data centers requires upfront capital investment. Once startup demand falls short of expectations, idle computing power may not be subleased at the anticipated price.